基于神经网络的甲状腺肿瘤复发风险评估模型-《中国医学物理学杂志》

文章信息/Info

Title:: A neural network-based model for predicting thyroid tumor recurrence risk

作者:: 罗爱静1; 3; 4; 5; 王哲轩1; 3; 4; 5; 谢文照2; 3; 4; 5; 胡德华3; 徐倩1; 3; 4; 5; 舒泳搏1; 3; 4; 5; 1.中南大学湘雅二医院，湖南长沙 410011；2. 中南大学湘雅三医院毕业后医学教育办公室，湖南长沙 410013；3.中南大学生命科学学院，湖南长沙 410013；4.医学信息研究湖南省普通高等学校重点实验室（中南大学），湖南长沙 410013；5.湖南省心血管智能医疗临床医学研究中心，湖南长沙 410011

Author(s):: LUO Aijing1; 3; 4; 5; WANG Zhexuan1; 3; 4; 5; XIE Wenzhao2; 3; 4; 5; HU Dehua3; XU Qian1; 3; 4; 5; SHU Yongbo1; 3; 4; 5; 1. The Second Xiangya Hospital of Central South University, Changsha 410011, China; 2. Post-Graduation Education Office, the ThirdXiangya Hospital of Central South University, Changsha 410013, China; 3. School of Life Sciences, Central South University,Changsha 410013, China; 4. Key Laboratory of Medical Information Research (Central South University), College of Hunan Province,Changsha 410013, China; 5. Clinical Research Center for Cardiovascular Intelligent Healthcare in Hunan Province, Changsha 410011,China

Keywords:: thyroid tumor; postoperative recurrence; machine learning; artificial neural network

摘要:: 目的：开发基于神经网络的甲状腺肿瘤患者术后复发预测的深度学习模型，并通过外部验证，为临床医生提供可靠的决策支持参考工具。方法：基于人工神经网络结构，使用SEER数据库筛选得到的甲状腺肿瘤数据作为训练集，并使用加利福尼亚大学尔湾分校（UCIrvine）公布的开源数据以及来自湖南省某大型三甲医院的100例患者数据进行外部验证。模型通过多项性能指标评估其预测复发的准确性与可靠性。结果：实验结果显示，该模型在复发预测中的表现优于Logistic模型。内部验证的具体结果为：准确率0.915 3，召回率0.981 8，精确率0.921 1，F1值为0.947 4；UCIrvine验证集上的具体结果为：准确率0.832 9，召回率0.945 5，精确率0.841 4，F1值为0.890 4，ROC_AUC为0.78；本地验证集的具体结果为：准确率0.870 0，召回率0.880 0，精确率0.862 7，F1值为0.871 3，ROC_AUC为0.80。结论：基于人工神经网络的预测模型在甲状腺肿瘤复发预测中表现出色，为临床医生提供有效的辅助决策工具，有助于优化术后治疗方案和提高患者预后管理。

Abstract:: Abstract: Objective To develop a neural network-based deep learning model for predicting postoperative recurrence inthyroid tumor patients and validate the model with external datasets for providing clinicians with a reliable decision supporttool. Methods An artificial neural network structure was adopted in the study, with thyroid tumor data from the SEERdatabase serving as the training set. External validation was conducted with open-source data from the University ofCalifornia, Irvine (UCIrvine), and the data from 100 patients at a general tertiary hospital in Hunan province. The model’saccuracy and reliability in predicting recurrence were evaluated through multiple performance metrics. Results Experimentalresults showed that the model outperformed Logistic model in recurrence prediction, with accuracy, recall rate, precision andF1 score reaching 0.915 3, 0.981 8, 0.921 1 and 0.947 4 in internal validation. Moreover, the model achieved accuracies,recall rates, precisions, F1 scores and ROC_AUC values of 0.832 9, 0.945 5, 0.841 4, 0.890 4 and 0.78 on the UCIrvinevalidation set, while 0.870 0, 0.880 0, 0.862 7, 0.871 3 and 0.80 on the local validation set. Conclusion This neural networkbased predictive model exhibits excellent performance in thyroid tumor recurrence prediction, providing clinicians with avaluable decision support tool that can help optimize postoperative treatment plans and improve patient prognosismanagement.